(Author here) Haha that is a great point. I was trying to come up with a term that described my personal workflow and specifically felt different than vibe coding (because it's geared towards how professional programmers can use agents). Very open to alternative terms!
This type of coding has been extremely helpful to me in the past few weeks. I’m on parental leave, but also a co-owner of a small company and can’t completely log off.
I can one handed spec out changes, AI does its thing, and then I review and refine it whenever my kid is asleep for 20 minutes. Or if I’m super tired I’m able to explain changes with horrible english and get results. At the same time, I am following a source control and code review process that I’ve used in large teams. I’ve even been leaving comments on PRs where AI contributes and I’m the only dev in the codebase.
I wouldn’t call this vibe coding— however vibe coding could be a subset of this type of work. I think async coding is a good description, but bad because of what it means as a software concept (which is mentioned). Maybe AI-delegation?
I want to understand the distinction you're making against vibe coding.
In vibe coding, the developer specifies only functional requirements (what the software must do) and non-functional requirements (the qualities it must have, like performance, scalability, or security). The AI delivers a complete implementation, and the developer reviews it solely against those behaviors and qualities. Any corrections are given again only in terms of requirements, never code, and the cycle repeats until the software aligns.
But you're trying to coin a term for the following?
In ??? coding, the developer specifies code changes that must be made, such as adding a feature, modifying an existing function, or removing unused logic. The AI delivers the complete set of changes to the codebase, and the developer reviews it at the code level. Any corrections are given again as updates to the code, and the cycle repeats until the code aligns.
Did I understand it right?
If so, I've most seen the latter be called AI pair-programming or AI-assisted coding. And I'd agree with the other commenters, please DO NOT call it async programming (even if you add async AI it's too confusing).
> In ??? coding, the developer specifies code changes that must be made, such as adding a feature, modifying an existing function, or removing unused logic. The AI delivers the complete set of changes to the codebase, and the developer reviews it at the code level. Any corrections are given again as updates to the code, and the cycle repeats until the code aligns.
Yes
> If so, I've most seen the latter be called AI pair-programming or AI-assisted coding.
I specifically considered both terms and am not a fan
* "pair-programming" is something that involves two people paying attention while writing code, and in this case, i'm not looking at the screen while the AI system writes code
* "AI-assisted coding" is generally anchored to copilots/IDE style agents where people are actively writing code, and an AI assists them.
I totally hear you on conflating async. However, I think the appropriate term would clearly indicate that this happens without actively watching the AI write code. Unfortunately I think other terms like "background" may also be confusing for similar reasons.
Agentic already implies what you want to imply. Maybe just call it agentic coding?
I feel that be the most clear. Agentic coding implies any workflow using AI agents. Which mean it's always the same agentic coding loop:
1. Prompt...
2. Wait or go do something else while agents make edits...
3. Come back to review the result
4. Go to 1
> is generally anchored to copilots/IDE style agents where people are actively writing code
I don't know when you last used these, but they're all agentic now. The workflow is exactly the same, you don't write code and auto-complete suggestions, you prompt and they go and make multiple edits to multiple files and can take upwards of 10 minutes, once done they show you a diff (or you can trust) and you're free to review/test or not, and prompt some more.
Edit: Or what the other commenter said: "prompt driven coding", that could be a good term as well.
I think it's still vibe coding. In practice any AI-driven process where you tell the AI what you want and it writes the code is considered "vibe coding."
No, it isn’t. There is a vast amount of difference between “vibe coding” and prompt driven development.
I’m giving the benefit of the doubt to the author here that it’s very unlikely they consider their example to be an actual representative scenario.
Vibe coding is to allow the AI to make the majority of the decisions. What the author describes is more like a highly complex autocomplete; you establish the fairly detailed outline of what is needed, often using tools/servers/etc tailored to use cases, and expect the AI to design an implementation that is in-line with the human-made decisions that preceded it, which is why I draw the comparison to autocomplete. Vibe coding is more like paying the kid next door to write your school essay…comparatively.
I used vibe coding to build a UI prototype of workflow. I used mockup images as the basis of the layout and let the agent use Redis as the persistence layer. I know it will be throw away and don't care how it is working underneath as long as it can demonstrate the flow I want.
I have also used prompt driven development to allow the agent to code something I expect to turn into a longer term product. I do more review of the code, ensure it is meeting all standards to development I would expect of myself or any other developer.
There are certainly differing degrees of the two types of development.
I did this early in my career as a product owner with an offshore team in India... Write feedback/specs, send them over at end of day US time. Have a fresh build ready for review by start of business.
Worked amazingly when it worked. Really stretched things out when the devs misunderstood us or got confused by our lack of clarity and we had to find time for a call... Also eventually there got to be some gnarly technical debt and things really slowed down.
I think it can only work if the product owner literally owns the product as in has FULL decision making power about what goes or doesn't go etc. it doesn't work when a product manager is a glorified in between guy, dictating the wishes of the CEO through a game of telephone from the management.
You’ll have to be more specific about what you mean by “product owner”, because that’s a very nebulous job title. For example, how technical is this product owner? Are they assumed to “just know” that they’re asking for an overly complex, expensive technical solution?
This would be incredibly useful for software development, period. A 10x factor, all by itself. Yet it happens infrequently, or, at best, in significantly limited ways.
The main problem, I think, is that it assumes you already know what you want at the start, and, implicitly, that what you want actually makes some real sense.
I guess maybe the context is cranking out REST endpoints or some other constrained detail of a larger thing. Then, sure.
Figuring out what you want is the hard part about programming. I think that's where AI augmentation will really shine, because it lowers the time between iterations and experiments.
That said, this article is basically describing being a product owner.
(Author here) I can certainly appreciate having an alternate perspective, but I think it's unfair to say it's DOA. I've personally used this workflow for the last 6 months and shipped a lot of features into our product, including the lowest levels of infra all the way to UI code. I definitely think there is a lot to improve. But it works, at least for me :)
> is that it assumes you already know what you want at the start, and, implicitly, that what you want actually makes some real sense.
My experience is different. I find that AI-powered coding agents drop the barriers to experimentation drastically, so that ... yes if I don't know what I Want, I can go try things very easily, and learn. Exploration just got soooo much cheaper. Now that may be a different interaction that what is described in this blog post. The exploration may be a precursor to what is happening in this blog post. But once I'm done exploring I can define the problem and ask for solutions.
If it's DOA you'd better tell everyone who is currently doing this, that they're not really doing this.
This works until you get to the point that your actual programming skills atrophy due to lack of use.
Face it, the only reason you can do a decent review is because of years of hard won lessons, not because you have years of reading code without writing any.
1. Learn how to describe what you want in an unambiguous dialect of natural language.
2. Submit it to a program that takes a long time to transform that input into a computer language.
3. Review the output for errors.
Sounds like we’ve reinvented compilers. Except they’re really bad and they take forever. Most people don’t have to review the assembly language / bytecode output of their compilers, because we expect them to actually work.
Product managers never get that right though. In practice it always falls back on the developer to understand the problem and fill in the missing pieces.
In many cases it falls on the developer to talk the PM out of the bad idea and then into a better solution. Agents aren’t equipped to do any of that.
For any non trivial problem, a PM with the same problem and 2 different dev teams will produce a drastically different solutions 99 times out of 100.
Dev teams are much less non-deterministic than LLMs. If you ask the same dev team to build the same product multiple times they’ll eventually converge on the producing the same product.
The 2nd time it will likely be pretty different because they’ll use what they learned to build it better. The 3rd time will be better still, but each time after that it will essentially be the same product.
An LLM will never converge. It definitely won’t learn from each subsequent iteration.
Human devs are also a lot more resilient to slight changes in requirements and wording. A slight change in language that wouldn’t impact a human at all will cause an LLM to produce completely different output.
An LLM within the right context/environment can also converge: just like with humans you need to provide guidelines, rules, and protocols to instruct how to implement something. Just like with humans I’ve used the approach you describe: generate something one until it works the way you want it, then ask it so document insights, patterns and rules, and for the next project instruct it to follow the rules you persisted. Will result in more or less the same project.
Humans are very non deterministic: if you ask me to solve a problem today, the solution will be different from last week, last year or 10 years ago. We’ve learnt to deal with it, and we can also control the non-determinism of LLMs.
And humans are also very prone to hallucinations: remember those 3000+ gods that we’ve created to explain the world, or those many religions that are completely incompatible? Even if some are true, most of them must be hallucinations just by being incompatible to the others.
That only works with very small projects to the point where the specification document is a very large percentage of the total code.
If you are very experienced, you won’t solve the problem differently day to day. You probably would with a 10 year difference, but you won’t ever be running the next model 10 years out (even if the technology matures), so there’s no point in doing that comparison. Solving the same problem with the same constraints in radically different ways day to day comes from inexperience (unless you’re exploring and doing it on purpose).
Calling what LLMs do hallucinations and comparing it to human mythology is stretching the analogy into absurdity.
If you work at "product manager" level, that would be vibe coding, as you prompt for functional and non-functional changes and you review the behavior and characteristics of the program, not the code generated.
I believe the author was trying to specifically distinguish their workflow from that, in that they are prompting for changes to the code in terms of the code itself, and reviewing the code that is generated (maybe along with also mentioning the functionality and testing it).
What I described is precisely the reception of early compilers. How is the LLM different? It’s slower? Its input looks more like natural language? Its output is less reliable? It runs on somebody else’s computer? What’s the essential difference between these two technologies that transform one text into another?
Is it the work of a product manager? I believe the latter only specify features and business rules (and maybe some other specifications like UX and performance). But no technical details at all. That would be like an architect reviewing the brand of nails used in an house framing.
I think this is better than many current coding interview methods. Assuming you have an agent setup to not give the interviewee the answer directly.
Of course there are times when you need someone extremely skilled at a particular language. But from my experience I would MUCH prefer to see how someone builds out a problem in natural language and have guarantees to its success. I’ve been in too many interviews where candidates trip over syntax, pick the wrong language, or are just not good at memorization and don’t want to look dumb looking things up. I usually prefer paired programming interviews where I cater my assistance to expectations of the position. AI can essentially do that for us.
Yeah research says the interview process should match the day to day expectations as closely as possible, even to a trial day/week/month. All these leetcode and tricky puzzles are very low on signal. They don't tell you how a person will do on the job at all, not to mention they're bad for women and minorities.
I have a static analysis and refactoring tool that does wonders to identify duplication and poor architecture patterns and provide a roadmap for agents to fix the issues. It's like magic, just point it at your codebase then tell the agent to grind away at the output (making sure to come up for air and rerun tests regularly) and it'll go for hours.
Official release is tomorrow, I'm just doing final release prep cleanup and getting the product page in order, but the crate/brew formula is in decent shape, just missing some features I'll be shipping soon. https://github.com/sibyllinesoft/valknut if you want to jump the line though.
A quick pass through this repository validates that there is a lot of tech debt in this code too. For a very simple example, consider that the analysis code is doing string searches to attempt to detect various syntatic constructs. https://github.com/sibyllinesoft/valknut/blob/aaf2b818a97b8d...
I realize, my marketing page even puts it front and center (I use the tool to analyze the tool, yay dogfood). I'm compute limited at the moment and I don't have cycles to burn refactoring this code base since it's pretty close to feature complete right now, I need to put all my compute towards finishing up development of a few projects I'm trying to ship this week, and some exhaustive bechmark matrixes.
Regarding that string search, you really have to fight Claude to get it to use tree sitter consistently, I have to do a search through my codebase to build an audit list for this stuff.
You're right, reviews aren't the way forward. We don't do code reviews on compiler output (unless you're writing a compiler). The way forward is strong static and analytic guardrails and stochastic error correction (multiple solutions proposed with LLM as a judge before implementation, multiple code review agents with different personas that have been prompted to be strict/adversarial but not nit-pick) with robust test suites that have also been through multiple passes of audits and red-teaming by agents. You should rarely have to look at the code, it should be a significant escalation event like when you need to coordinate with Apple due to XCode bugs.
Unless you are writing some shitty code for a random product that will be used for some demo then trashed, code can be resumed to a simple thing:
Code is a way to move ideas into the real world through a keyboard
So, reading that the future is using a random machine with an averaged output (by design), but that this output of average quality will be good enough because the same random machine will generate tests of the same quality : this is ridiculous
Tests are probably the thing you should never build randomly, you should put a lot of thoughts in them: do they make sense ? Do your code make sense ? With tests, you are forced to use your own code, sometimes as your users will
Writing tests is a good way to force yourself to be empathic with your users
People that are coding through IA are the equivalent of the pre-2015 area system administrators that renewed TLS certificates manually. They are people that can (and are replacing themselves) with bash scripts. I don't miss them and I won't miss this new kind.
I actually have a bayesian stochastic process model for LLM codegen that incorporates the noisy channel coding theorem, it turns out that just like noisy communications channels can be encoded to give arbitrarily low error rate communication, LLM agents workflows can be coded to give arbitrarily low final error rate output. The only limitation on this is when model priors are highly mis-aligned with the work that needs to be done, in that case you need hard steering via additional context.
Creative outputs start with gemini (because of long context support, it can get longform stuff right), with successive refinement passes using claude for line/copy edits because (it's the least purple).
> You should rarely have to look at the code, it should be a significant escalation event
This is the bit I am having problems with: if you are rarely looking at the code, you will never have the skills to actually debug that significant escalation event.
For generative skills I agree, but for me the real change is in how I read and debug code. After reading so much AI-generated code with subtle mistakes, I can spot errors much quicker even in human-written code. And when I can't, that usually means the code needs a refactor.
I'd compare it to gym work: some exercises work best until they don't, and then you switch to a less effective exercise to get you out of your plateau. Same with code and AI. If you're already good (because of years of hard won lessons), it can push you that extra bit.
But yeah, default to the better exercise and just code yourself, at least on the project's core.
I mean that I've read so much AI generated code with subtle mistakes that my brain jumps straight to the likely failure point and I've noticed it generalizes. Even when I look at an OSS project I'm not super familiar with, I can usually spot the bugs faster then before. I'll edit my initial response for clarity.
(Author here) Personally, I try to combat this by synchronously working on 1 task and asynchronously working on others. I am not sure it's perfect, but it definitely helps me avoid atrophy.
I actually like writing code, it does get tedious I get that when you're making yet another component. I don't feel joy when you just will a bunch of code into existence with words. It's like actively participating in development when typing. Which yeah people use libraries/frameworks/boilerplate.
My dream is to not be employed in software and do it for fun (or work on something I actually care about)
Even if I wrote some piece of crap, it is my piece of crap
Some people will probably call you a luddite, but don't listen to them. There's nothing wrong with taking joy in the craft, with learning and exploring and creating. That's what hacker culture used to be about.
Unfortunately, you won't be able to get a job in software with anything but AI skills, since humans no longer write software in the industry. People will look at you the way they used to look at anyone who wrote their own HTML or Javascript without frameworks and Typescript, like you must drive your car to work with your feet.
It was funny I was handed this project to work on and I skimmed the ReadMe, there were a lot of readmes in the code like how to use pipenv or whatever basic stuff... at first I was like "nice job with the docs" but then I later realized it was a vibe coded project I felt like I was owned. It's funny
Also funny how much time was wasted since it had random code in it that was not removed (not working old code vs. current working new code) that's not to blame on the AI part but yeah.
I have a job now in the industry it's funny I work with AI eg. AWS Bedrock/Knowledgebases/Agents... RAG/LLM AI.
The AI I want to work with is vision/ML (robotics) but don't have the background for that (I do it as a hobby instead).
I'm feeling the effect of vibe coding now, where the 2nd leader in our team was only recently a developer but uses ChatGPT/windsurf to code for him which enables him to work on random topics like OpenSearch one day Airflow the next... idk I get I'm the one being left behind by not doing it too but I also want to really learn/understand something. You can do that with an AI-assisted thing but yeah... idk I don't want to that's what I'm saying, I will get out eventually once I've saved enough money.
My learning process for a while has been watching YT crash courses/reading the docs/finding articles...
The project I mentioned above there was literally a prompt in the repo "Write me an event-driven app with this architecture..."
The 2nd leader I mentioned above is a code at work/not at home type of person which is fine but yeah. I'm not that person, I like to actually code/make stuff outside of work. It's not just about getting a task done/shipping some code for me. But I guess that's what a business is, churn out something.
Idk there's some validity there isn't there... "I've been a developer for 10 years, then a guy with 2 years comes in vibe coding stuff" is the leader. Which I'm past it, I don't do office politics anymore, I've got a six-fig job, no need to climb, I'm coasting. Debt is really the only problem I have.
Those of us who worked in hardware, or are old programmers will find this familiar. Chip/board routing jobs that took days to complete. Product build/test jobs that took hours to run.
See also that movie with Johnny Depp where AI takes over the world.
Hi everyone, thanks for the spirited debate! I think there are some great points in the discussion so far. Some thoughts:
* "This didn't work for offshoring, why will it work all of a sudden?" I think there are good lessons to draw from offshoring around problem definition and what-not but the key difference is the iteration speed. Agents allow you to review stuff much faster, and you can look at smaller pieces of incremental work.
* "I thought this would be about async primitives in python, etc" Whoops sorry, I can understand how the name is confusing/ambiguous! The use of "async" here refers to the fact that I'm not synchronously looking at an IDE while writing code all the time.
* "You can only do this because you used to handwrite code". I don't think this workflow is a replacement for handwriting code. I still love doing that. This workflow just helps me do more.
I think this does not bode well for you; be honest with yourself - if you're making simple mistakes like using the term "Async Programming" to refer to something new, your prompting and/or code reviewing is probably not going all that well.
Sure, it can look good now, when there's no legacy, but if you ever move into having to maintain that code you're going to be in a tough spot.
I do think that AI will work well compared to the low end of offshoring, where to get good results you need people who could do the work themselves tightly involved. AI will give you slop code faster and cheaper, and that is sometimes enough.
The question is how it compares to the medium level of offshoring. Near term I think that at comparable cost ($100s of dollars per week), it'll give faster results at an acceptable tradeoff in quality for most uses. I don't think most companies want to spend thousands of dollars a month on developer tools per developer though... even though they often do.
This kind of workflow really doesn't appeal to me in the slightest. Maybe it works for some people, but it just seems to drain all the pleasure out of programming. For me, at least, solving the little problems are like little satisfying puzzles which makes it easier to maintain motivation.
> My workflow has changed since adopting async programming. I now work on four or five tasks simultaneously: one complex problem synchronously and three or four in the background.
Someone, please, try to convince me why this is a positive thing.
Perhaps the background tasks are ones with obvious solutions? I use Claude code to write boilerplate terraform for me as I know what it should look like and I often don't want to write it myself as it is not novel or exciting. I could see the argument that I batch a couple of these out to AI, focus on something more complicated, and then come back later to essentially code review them or make necessary adjustments.
I actually enjoy writing code... most of the time. I find myself turning to AI to write code I have an aversion to writing, not as a substitute for my own practice, but to get code that I would not have written in the first place. Like benchmarks, bash scripts, dashboards, unit tests, etc.
I can live without these things, but they're nice to have without expending the effort to figure out all the boilerplate necessary for solving very simple problems at their core. Sometimes AI can't get all the way to a solution, but usually it sets up enough of the boilerplate that only the fun part remains, and that's easy enough to do.
I don’t know anything about the competency of the AI boosters online.
But among people I’ve worked with whose capabilities I can judge, the competent programmers are not building things like this. Among my own sample size there is a near perfect negative correlation between AI use and competency.
Sounds great in principle but I have been trained to value individuals and interactions over processes and tools and working software over comprehensive documentation.
I don't know why we need a term like "Async AI programming." this is literally what you would do if you were a Tech Lead directing a team of other developers. You define what you want and hand it to one of your devs.
This is just being a TL. the agent is an assistant or a member of the team. I don't know why we need to call it "Async AI programming", unless we want to shy away from or obscure the idea that the agent is actually performing the job a human used to perform.
I am teaching asynchronous programming in typescript to junior developpers.
And i find really tricky to tell them that async and await do MAJOR magic behind their back to make their code readable as synchronous code.
And then, I need to detail very precisely what "Promise.all()" (and "return") really mean in the context of async/await. Which is something that (I feel) could have been abstracted away during the async/await syntax definition, and make the full magic much more natural.
Async/await themselves are not that much magic really, it's a bit of syntactic sugar over promise chains. Of course, understanding promises is its own bag.
To elaborate a bit, telling them that you should not "aList.foreach(asyncMethod)", but you'd better do "Promise.all(aList.map(asyncMethod))" is NOT very easy for them.
I think there is a confusion here between Coding and Programming. I think what is described here as "Async Programming" is just programming the way it should be which is different than coding. This is what Leslie Lamport pointed out a while back [1] and recently [2]. According to him programming has 3 stages:
1- Define what task the program should perform
2- Define how the program should do it
3- Writing the code that does it.
Most SWEs usually skip to step 3 instead of going through 1 and 2 without giving it much thought, and implement their code iteratively. I think Step 3 also includes testing, review, etc.
With AI developers are forced to think about the functionality and the specs of their code to pass it to AI to do the job and can no longer just jump to step 3. For delegating to other devs, the same process is required, senior engineers usually create design docs and pass it to junior engineers.
IMO automated verification and code reviews are already part of many developers workflows, so it's nothing new.
I get the point of the article though, that there are new requirements for programming and things are different in terms of how folks approach programming. So I do not agree that the method is new or should be called "async", it's the same method with brand new tools.
Effective async programming specs read like technical documentation
The thing I like least about software engineering will now become the primary task. It's a sad future for me, but maybe a great one for some different personality type.
Indeed! Why not just call it "asynchronous software development" or something similar? "asynchronous programming" is a bad choice, partly because it will be un-googleable.
They don't really as once the spec gets detailed enough it becomes so large and unwieldy that nobody with any actual power reads the things.
An executive at a large company once told me about something where a spec had been written and reviewed by all relevant stakeholders: "That may be what I asked for, but its not what I want."
Before I read the article I thought this meant programming with "async".
Just call it Agent-based programming or somesuch, otherwise it's really confusing!
(Author here) Haha that is a great point. I was trying to come up with a term that described my personal workflow and specifically felt different than vibe coding (because it's geared towards how professional programmers can use agents). Very open to alternative terms!
This type of coding has been extremely helpful to me in the past few weeks. I’m on parental leave, but also a co-owner of a small company and can’t completely log off.
I can one handed spec out changes, AI does its thing, and then I review and refine it whenever my kid is asleep for 20 minutes. Or if I’m super tired I’m able to explain changes with horrible english and get results. At the same time, I am following a source control and code review process that I’ve used in large teams. I’ve even been leaving comments on PRs where AI contributes and I’m the only dev in the codebase.
I wouldn’t call this vibe coding— however vibe coding could be a subset of this type of work. I think async coding is a good description, but bad because of what it means as a software concept (which is mentioned). Maybe AI-delegation?
There is already a term for it! It's called "Ralph coding": https://ghuntley.com/ralph/
I want to understand the distinction you're making against vibe coding.
In vibe coding, the developer specifies only functional requirements (what the software must do) and non-functional requirements (the qualities it must have, like performance, scalability, or security). The AI delivers a complete implementation, and the developer reviews it solely against those behaviors and qualities. Any corrections are given again only in terms of requirements, never code, and the cycle repeats until the software aligns.
But you're trying to coin a term for the following?
In ??? coding, the developer specifies code changes that must be made, such as adding a feature, modifying an existing function, or removing unused logic. The AI delivers the complete set of changes to the codebase, and the developer reviews it at the code level. Any corrections are given again as updates to the code, and the cycle repeats until the code aligns.
Did I understand it right?
If so, I've most seen the latter be called AI pair-programming or AI-assisted coding. And I'd agree with the other commenters, please DO NOT call it async programming (even if you add async AI it's too confusing).
> In ??? coding, the developer specifies code changes that must be made, such as adding a feature, modifying an existing function, or removing unused logic. The AI delivers the complete set of changes to the codebase, and the developer reviews it at the code level. Any corrections are given again as updates to the code, and the cycle repeats until the code aligns.
Yes
> If so, I've most seen the latter be called AI pair-programming or AI-assisted coding.
I specifically considered both terms and am not a fan * "pair-programming" is something that involves two people paying attention while writing code, and in this case, i'm not looking at the screen while the AI system writes code * "AI-assisted coding" is generally anchored to copilots/IDE style agents where people are actively writing code, and an AI assists them.
I totally hear you on conflating async. However, I think the appropriate term would clearly indicate that this happens without actively watching the AI write code. Unfortunately I think other terms like "background" may also be confusing for similar reasons.
Agentic already implies what you want to imply. Maybe just call it agentic coding?
I feel that be the most clear. Agentic coding implies any workflow using AI agents. Which mean it's always the same agentic coding loop:
> is generally anchored to copilots/IDE style agents where people are actively writing codeI don't know when you last used these, but they're all agentic now. The workflow is exactly the same, you don't write code and auto-complete suggestions, you prompt and they go and make multiple edits to multiple files and can take upwards of 10 minutes, once done they show you a diff (or you can trust) and you're free to review/test or not, and prompt some more.
Edit: Or what the other commenter said: "prompt driven coding", that could be a good term as well.
I think it's still vibe coding. In practice any AI-driven process where you tell the AI what you want and it writes the code is considered "vibe coding."
No, it isn’t. There is a vast amount of difference between “vibe coding” and prompt driven development.
I’m giving the benefit of the doubt to the author here that it’s very unlikely they consider their example to be an actual representative scenario.
Vibe coding is to allow the AI to make the majority of the decisions. What the author describes is more like a highly complex autocomplete; you establish the fairly detailed outline of what is needed, often using tools/servers/etc tailored to use cases, and expect the AI to design an implementation that is in-line with the human-made decisions that preceded it, which is why I draw the comparison to autocomplete. Vibe coding is more like paying the kid next door to write your school essay…comparatively.
Agreed. There is a difference.
I used vibe coding to build a UI prototype of workflow. I used mockup images as the basis of the layout and let the agent use Redis as the persistence layer. I know it will be throw away and don't care how it is working underneath as long as it can demonstrate the flow I want.
I have also used prompt driven development to allow the agent to code something I expect to turn into a longer term product. I do more review of the code, ensure it is meeting all standards to development I would expect of myself or any other developer.
There are certainly differing degrees of the two types of development.
Same! I was hoping this would have some insights into pitfalls or the like with javascript promises or python async, but alas no such luck.
Exactly. I think the traditional meaning of “asynchronous programming” was coined first. So, let’s stick with that.
Same here. I've read the author's braintrust.dev as "brain - Rust - Dev", so I was expecting a discussion on Rust Async development.
I did this early in my career as a product owner with an offshore team in India... Write feedback/specs, send them over at end of day US time. Have a fresh build ready for review by start of business.
Worked amazingly when it worked. Really stretched things out when the devs misunderstood us or got confused by our lack of clarity and we had to find time for a call... Also eventually there got to be some gnarly technical debt and things really slowed down.
I think it can only work if the product owner literally owns the product as in has FULL decision making power about what goes or doesn't go etc. it doesn't work when a product manager is a glorified in between guy, dictating the wishes of the CEO through a game of telephone from the management.
You’ll have to be more specific about what you mean by “product owner”, because that’s a very nebulous job title. For example, how technical is this product owner? Are they assumed to “just know” that they’re asking for an overly complex, expensive technical solution?
I'd guess they'd be assumed to "just know" to trust the developers working with them on that?
> it can only work if the product owner literally owns the product as in has FULL decision making power
This seems like a fairly rare situation in my experience.
It's not uncommon in the sort of solo-dev bootstrapped startup that is going wild for AI coding right now though.
I'm surprised there aren't more of these bubbling up, I can't name a single company like that.
Agreed. A glorified go between person is rarely going to succeed at delivering something good.
This vision of AI programming is DOA.
The first step is "define the problem clearly".
This would be incredibly useful for software development, period. A 10x factor, all by itself. Yet it happens infrequently, or, at best, in significantly limited ways.
The main problem, I think, is that it assumes you already know what you want at the start, and, implicitly, that what you want actually makes some real sense.
I guess maybe the context is cranking out REST endpoints or some other constrained detail of a larger thing. Then, sure.
I disagree with being detailed, many times I want to AI to think of things, half the time it comes up with something I wouldn't have that I like.
The thing I would add is to retry to prompt, don't tell it to fix a mistake. Rewind and change the prompt to tell It not to do that it did.
I agree there is a lot of value to have it do, what it considers, the obvious thing.
It is almost by definition what the average programmer would expect to find, so it's valuable as such.
But the moment you want to do something original, you need to keep high-level high-quality documentation somewhere.
Figuring out what you want is the hard part about programming. I think that's where AI augmentation will really shine, because it lowers the time between iterations and experiments.
That said, this article is basically describing being a product owner.
(Author here) I can certainly appreciate having an alternate perspective, but I think it's unfair to say it's DOA. I've personally used this workflow for the last 6 months and shipped a lot of features into our product, including the lowest levels of infra all the way to UI code. I definitely think there is a lot to improve. But it works, at least for me :)
> is that it assumes you already know what you want at the start, and, implicitly, that what you want actually makes some real sense.
My experience is different. I find that AI-powered coding agents drop the barriers to experimentation drastically, so that ... yes if I don't know what I Want, I can go try things very easily, and learn. Exploration just got soooo much cheaper. Now that may be a different interaction that what is described in this blog post. The exploration may be a precursor to what is happening in this blog post. But once I'm done exploring I can define the problem and ask for solutions.
If it's DOA you'd better tell everyone who is currently doing this, that they're not really doing this.
This works until you get to the point that your actual programming skills atrophy due to lack of use.
Face it, the only reason you can do a decent review is because of years of hard won lessons, not because you have years of reading code without writing any.
What the article describes is:
1. Learn how to describe what you want in an unambiguous dialect of natural language.
2. Submit it to a program that takes a long time to transform that input into a computer language.
3. Review the output for errors.
Sounds like we’ve reinvented compilers. Except they’re really bad and they take forever. Most people don’t have to review the assembly language / bytecode output of their compilers, because we expect them to actually work.
No, it sounds like the work of a product manager, you’re just working with agents rather than with developers.
Product managers never get that right though. In practice it always falls back on the developer to understand the problem and fill in the missing pieces.
In many cases it falls on the developer to talk the PM out of the bad idea and then into a better solution. Agents aren’t equipped to do any of that.
For any non trivial problem, a PM with the same problem and 2 different dev teams will produce a drastically different solutions 99 times out of 100.
Agree with the last bit, dev teams are even more non-deterministic than LLMs.
Dev teams are much less non-deterministic than LLMs. If you ask the same dev team to build the same product multiple times they’ll eventually converge on the producing the same product.
The 2nd time it will likely be pretty different because they’ll use what they learned to build it better. The 3rd time will be better still, but each time after that it will essentially be the same product.
An LLM will never converge. It definitely won’t learn from each subsequent iteration.
Human devs are also a lot more resilient to slight changes in requirements and wording. A slight change in language that wouldn’t impact a human at all will cause an LLM to produce completely different output.
An LLM within the right context/environment can also converge: just like with humans you need to provide guidelines, rules, and protocols to instruct how to implement something. Just like with humans I’ve used the approach you describe: generate something one until it works the way you want it, then ask it so document insights, patterns and rules, and for the next project instruct it to follow the rules you persisted. Will result in more or less the same project.
Humans are very non deterministic: if you ask me to solve a problem today, the solution will be different from last week, last year or 10 years ago. We’ve learnt to deal with it, and we can also control the non-determinism of LLMs.
And humans are also very prone to hallucinations: remember those 3000+ gods that we’ve created to explain the world, or those many religions that are completely incompatible? Even if some are true, most of them must be hallucinations just by being incompatible to the others.
That only works with very small projects to the point where the specification document is a very large percentage of the total code.
If you are very experienced, you won’t solve the problem differently day to day. You probably would with a 10 year difference, but you won’t ever be running the next model 10 years out (even if the technology matures), so there’s no point in doing that comparison. Solving the same problem with the same constraints in radically different ways day to day comes from inexperience (unless you’re exploring and doing it on purpose).
Calling what LLMs do hallucinations and comparing it to human mythology is stretching the analogy into absurdity.
If you work at "product manager" level, that would be vibe coding, as you prompt for functional and non-functional changes and you review the behavior and characteristics of the program, not the code generated.
I believe the author was trying to specifically distinguish their workflow from that, in that they are prompting for changes to the code in terms of the code itself, and reviewing the code that is generated (maybe along with also mentioning the functionality and testing it).
What I described is precisely the reception of early compilers. How is the LLM different? It’s slower? Its input looks more like natural language? Its output is less reliable? It runs on somebody else’s computer? What’s the essential difference between these two technologies that transform one text into another?
Is it the work of a product manager? I believe the latter only specify features and business rules (and maybe some other specifications like UX and performance). But no technical details at all. That would be like an architect reviewing the brand of nails used in an house framing.
Tech Lead, not PM. (in my experience)
Coding interview of the future: "Show us how you would prompt this binary sort."
I think this is better than many current coding interview methods. Assuming you have an agent setup to not give the interviewee the answer directly.
Of course there are times when you need someone extremely skilled at a particular language. But from my experience I would MUCH prefer to see how someone builds out a problem in natural language and have guarantees to its success. I’ve been in too many interviews where candidates trip over syntax, pick the wrong language, or are just not good at memorization and don’t want to look dumb looking things up. I usually prefer paired programming interviews where I cater my assistance to expectations of the position. AI can essentially do that for us.
Yeah research says the interview process should match the day to day expectations as closely as possible, even to a trial day/week/month. All these leetcode and tricky puzzles are very low on signal. They don't tell you how a person will do on the job at all, not to mention they're bad for women and minorities.
preach. Take home assignments are another example. It gives huge bias towards those with free time.
My understanding is it's already here [1]
[1] https://news.ycombinator.com/item?id=44723289
not a joke.
Also, the future you are referring to is... like... 6 weeks from now.
Agreed.
> Hand it off. Delegate the implementation to an AI agent, a teammate, or even your future self with comprehensive notes.
The AI agent just feels like a way to create tech debt on a massive scale while not being able to identify it as tech debt.
I have a static analysis and refactoring tool that does wonders to identify duplication and poor architecture patterns and provide a roadmap for agents to fix the issues. It's like magic, just point it at your codebase then tell the agent to grind away at the output (making sure to come up for air and rerun tests regularly) and it'll go for hours.
What's it called?
Official release is tomorrow, I'm just doing final release prep cleanup and getting the product page in order, but the crate/brew formula is in decent shape, just missing some features I'll be shipping soon. https://github.com/sibyllinesoft/valknut if you want to jump the line though.
A quick pass through this repository validates that there is a lot of tech debt in this code too. For a very simple example, consider that the analysis code is doing string searches to attempt to detect various syntatic constructs. https://github.com/sibyllinesoft/valknut/blob/aaf2b818a97b8d...
I realize, my marketing page even puts it front and center (I use the tool to analyze the tool, yay dogfood). I'm compute limited at the moment and I don't have cycles to burn refactoring this code base since it's pretty close to feature complete right now, I need to put all my compute towards finishing up development of a few projects I'm trying to ship this week, and some exhaustive bechmark matrixes.
Regarding that string search, you really have to fight Claude to get it to use tree sitter consistently, I have to do a search through my codebase to build an audit list for this stuff.
This is what a lot of business leaders miss.
The benefits you might gain from LLMs is that you are able to discern good output from bad.
Once that's lost, the output of these tools becomes a complete gamble.
The business leaders already can't discern good from bad.
You're right, reviews aren't the way forward. We don't do code reviews on compiler output (unless you're writing a compiler). The way forward is strong static and analytic guardrails and stochastic error correction (multiple solutions proposed with LLM as a judge before implementation, multiple code review agents with different personas that have been prompted to be strict/adversarial but not nit-pick) with robust test suites that have also been through multiple passes of audits and red-teaming by agents. You should rarely have to look at the code, it should be a significant escalation event like when you need to coordinate with Apple due to XCode bugs.
Static and analytic guardrails ??
Unless you are writing some shitty code for a random product that will be used for some demo then trashed, code can be resumed to a simple thing:
So, reading that the future is using a random machine with an averaged output (by design), but that this output of average quality will be good enough because the same random machine will generate tests of the same quality : this is ridiculousTests are probably the thing you should never build randomly, you should put a lot of thoughts in them: do they make sense ? Do your code make sense ? With tests, you are forced to use your own code, sometimes as your users will
Writing tests is a good way to force yourself to be empathic with your users
People that are coding through IA are the equivalent of the pre-2015 area system administrators that renewed TLS certificates manually. They are people that can (and are replacing themselves) with bash scripts. I don't miss them and I won't miss this new kind.
I actually have a bayesian stochastic process model for LLM codegen that incorporates the noisy channel coding theorem, it turns out that just like noisy communications channels can be encoded to give arbitrarily low error rate communication, LLM agents workflows can be coded to give arbitrarily low final error rate output. The only limitation on this is when model priors are highly mis-aligned with the work that needs to be done, in that case you need hard steering via additional context.
Which model gives you creatives outputs ?
Creative outputs start with gemini (because of long context support, it can get longform stuff right), with successive refinement passes using claude for line/copy edits because (it's the least purple).
> You should rarely have to look at the code, it should be a significant escalation event
This is the bit I am having problems with: if you are rarely looking at the code, you will never have the skills to actually debug that significant escalation event.
good fucking luck writing adequate test suites for qualitative business logic
if it's even possible it will be more work than writing the code manually
For generative skills I agree, but for me the real change is in how I read and debug code. After reading so much AI-generated code with subtle mistakes, I can spot errors much quicker even in human-written code. And when I can't, that usually means the code needs a refactor.
I'd compare it to gym work: some exercises work best until they don't, and then you switch to a less effective exercise to get you out of your plateau. Same with code and AI. If you're already good (because of years of hard won lessons), it can push you that extra bit.
But yeah, default to the better exercise and just code yourself, at least on the project's core.
What do you mean you can spot errors much quicker?
I mean that I've read so much AI generated code with subtle mistakes that my brain jumps straight to the likely failure point and I've noticed it generalizes. Even when I look at an OSS project I'm not super familiar with, I can usually spot the bugs faster then before. I'll edit my initial response for clarity.
> subtle mistakes that my brain jumps straight to the likely failure ... I can usually spot the bugs faster then before
doubt intensifies
Doubt accepted. A spot-the-bug challenge on real OSS/prod code would be fun.
(Author here) Personally, I try to combat this by synchronously working on 1 task and asynchronously working on others. I am not sure it's perfect, but it definitely helps me avoid atrophy.
By synchronously working on 1 do you mean coding it with minimal AI?
Nice article by the way. I've found my workflow to be pretty much exactly the same using Claude code.
so... normal team lead -> manager pipeline?
Idk if I'm a luddite or what
I actually like writing code, it does get tedious I get that when you're making yet another component. I don't feel joy when you just will a bunch of code into existence with words. It's like actively participating in development when typing. Which yeah people use libraries/frameworks/boilerplate.
My dream is to not be employed in software and do it for fun (or work on something I actually care about)
Even if I wrote some piece of crap, it is my piece of crap
Some people will probably call you a luddite, but don't listen to them. There's nothing wrong with taking joy in the craft, with learning and exploring and creating. That's what hacker culture used to be about.
Unfortunately, you won't be able to get a job in software with anything but AI skills, since humans no longer write software in the industry. People will look at you the way they used to look at anyone who wrote their own HTML or Javascript without frameworks and Typescript, like you must drive your car to work with your feet.
It was funny I was handed this project to work on and I skimmed the ReadMe, there were a lot of readmes in the code like how to use pipenv or whatever basic stuff... at first I was like "nice job with the docs" but then I later realized it was a vibe coded project I felt like I was owned. It's funny
Also funny how much time was wasted since it had random code in it that was not removed (not working old code vs. current working new code) that's not to blame on the AI part but yeah.
I have a job now in the industry it's funny I work with AI eg. AWS Bedrock/Knowledgebases/Agents... RAG/LLM AI.
The AI I want to work with is vision/ML (robotics) but don't have the background for that (I do it as a hobby instead).
I'm feeling the effect of vibe coding now, where the 2nd leader in our team was only recently a developer but uses ChatGPT/windsurf to code for him which enables him to work on random topics like OpenSearch one day Airflow the next... idk I get I'm the one being left behind by not doing it too but I also want to really learn/understand something. You can do that with an AI-assisted thing but yeah... idk I don't want to that's what I'm saying, I will get out eventually once I've saved enough money.
My learning process for a while has been watching YT crash courses/reading the docs/finding articles...
The project I mentioned above there was literally a prompt in the repo "Write me an event-driven app with this architecture..."
The 2nd leader I mentioned above is a code at work/not at home type of person which is fine but yeah. I'm not that person, I like to actually code/make stuff outside of work. It's not just about getting a task done/shipping some code for me. But I guess that's what a business is, churn out something.
Idk there's some validity there isn't there... "I've been a developer for 10 years, then a guy with 2 years comes in vibe coding stuff" is the leader. Which I'm past it, I don't do office politics anymore, I've got a six-fig job, no need to climb, I'm coasting. Debt is really the only problem I have.
That's not the kind of async programming I was expecting.
I was ready for a deep-dive into things like asyncio in python; where it came from and what problems it promised to solve!
Those of us who worked in hardware, or are old programmers will find this familiar. Chip/board routing jobs that took days to complete. Product build/test jobs that took hours to run.
See also that movie with Johnny Depp where AI takes over the world.
(Author here)
Hi everyone, thanks for the spirited debate! I think there are some great points in the discussion so far. Some thoughts:
* "This didn't work for offshoring, why will it work all of a sudden?" I think there are good lessons to draw from offshoring around problem definition and what-not but the key difference is the iteration speed. Agents allow you to review stuff much faster, and you can look at smaller pieces of incremental work.
* "I thought this would be about async primitives in python, etc" Whoops sorry, I can understand how the name is confusing/ambiguous! The use of "async" here refers to the fact that I'm not synchronously looking at an IDE while writing code all the time.
* "You can only do this because you used to handwrite code". I don't think this workflow is a replacement for handwriting code. I still love doing that. This workflow just helps me do more.
I think this does not bode well for you; be honest with yourself - if you're making simple mistakes like using the term "Async Programming" to refer to something new, your prompting and/or code reviewing is probably not going all that well.
Sure, it can look good now, when there's no legacy, but if you ever move into having to maintain that code you're going to be in a tough spot.
I do think that AI will work well compared to the low end of offshoring, where to get good results you need people who could do the work themselves tightly involved. AI will give you slop code faster and cheaper, and that is sometimes enough.
The question is how it compares to the medium level of offshoring. Near term I think that at comparable cost ($100s of dollars per week), it'll give faster results at an acceptable tradeoff in quality for most uses. I don't think most companies want to spend thousands of dollars a month on developer tools per developer though... even though they often do.
It's just a different workflow IMO. AI is effectively real-time, whereas offshoring, no matter the quality, is something you have to do in batches.
This kind of workflow really doesn't appeal to me in the slightest. Maybe it works for some people, but it just seems to drain all the pleasure out of programming. For me, at least, solving the little problems are like little satisfying puzzles which makes it easier to maintain motivation.
> My workflow has changed since adopting async programming. I now work on four or five tasks simultaneously: one complex problem synchronously and three or four in the background.
Someone, please, try to convince me why this is a positive thing.
Perhaps the background tasks are ones with obvious solutions? I use Claude code to write boilerplate terraform for me as I know what it should look like and I often don't want to write it myself as it is not novel or exciting. I could see the argument that I batch a couple of these out to AI, focus on something more complicated, and then come back later to essentially code review them or make necessary adjustments.
It takes my longer to thoroughly review code I didn't write, especially code written by a junior developer.
Why would I choose to slow myself down in the short term and allow my skills to atrophy in the long term (which will also slow me down)?
I actually enjoy writing code... most of the time. I find myself turning to AI to write code I have an aversion to writing, not as a substitute for my own practice, but to get code that I would not have written in the first place. Like benchmarks, bash scripts, dashboards, unit tests, etc.
I can live without these things, but they're nice to have without expending the effort to figure out all the boilerplate necessary for solving very simple problems at their core. Sometimes AI can't get all the way to a solution, but usually it sets up enough of the boilerplate that only the fun part remains, and that's easy enough to do.
That sounds reasonable and similar to how I use it.
Managing a team of interns isn't fun, and I have no idea why someone who is a competent developer would choose to do that to themselves.
I don’t know anything about the competency of the AI boosters online.
But among people I’ve worked with whose capabilities I can judge, the competent programmers are not building things like this. Among my own sample size there is a near perfect negative correlation between AI use and competency.
Sounds great in principle but I have been trained to value individuals and interactions over processes and tools and working software over comprehensive documentation.
I don't know why we need a term like "Async AI programming." this is literally what you would do if you were a Tech Lead directing a team of other developers. You define what you want and hand it to one of your devs.
This is just being a TL. the agent is an assistant or a member of the team. I don't know why we need to call it "Async AI programming", unless we want to shy away from or obscure the idea that the agent is actually performing the job a human used to perform.
I am teaching asynchronous programming in typescript to junior developpers. And i find really tricky to tell them that async and await do MAJOR magic behind their back to make their code readable as synchronous code.
And then, I need to detail very precisely what "Promise.all()" (and "return") really mean in the context of async/await. Which is something that (I feel) could have been abstracted away during the async/await syntax definition, and make the full magic much more natural.
Async/await themselves are not that much magic really, it's a bit of syntactic sugar over promise chains. Of course, understanding promises is its own bag.
ChatGPT explanation: https://chatgpt.com/share/68c30421-be3c-8011-8431-8f3385a654...
During my interviews, may be I should ask them to read and understand this:
https://developer.mozilla.org/en-US/docs/Web/JavaScript/Refe...
prior to any dev they plan to do in JS/TS.
PS: 10 bucks that none of them would stay.
That reminds me of my Unix guru of the 90s: "man pages ARE easy to read".
[spoil: "when you are already an expert of the tool detailled in it"]
Man are you going to be disappointed when you read the article.
Man, for the first time in HN, I am teased to actually read the article.
Update: oh my god, I read the article. And feel completely cheated!!!!
Note for my future self: continue to read only the HN comments
To elaborate a bit, telling them that you should not "aList.foreach(asyncMethod)", but you'd better do "Promise.all(aList.map(asyncMethod))" is NOT very easy for them.
The article felt too rosy as I read it and then it got to the part where it’s just an ad for the author’s company.
I think there is a confusion here between Coding and Programming. I think what is described here as "Async Programming" is just programming the way it should be which is different than coding. This is what Leslie Lamport pointed out a while back [1] and recently [2]. According to him programming has 3 stages:
Most SWEs usually skip to step 3 instead of going through 1 and 2 without giving it much thought, and implement their code iteratively. I think Step 3 also includes testing, review, etc.With AI developers are forced to think about the functionality and the specs of their code to pass it to AI to do the job and can no longer just jump to step 3. For delegating to other devs, the same process is required, senior engineers usually create design docs and pass it to junior engineers.
IMO automated verification and code reviews are already part of many developers workflows, so it's nothing new.
I get the point of the article though, that there are new requirements for programming and things are different in terms of how folks approach programming. So I do not agree that the method is new or should be called "async", it's the same method with brand new tools.
Effective async programming specs read like technical documentation
The thing I like least about software engineering will now become the primary task. It's a sad future for me, but maybe a great one for some different personality type.
Oh, async?
> This version of "async programming" is different from the classic definition. It's about how developers approach building software.
Oh async=you wait until it is done. How interesting.
Where are the agents running? Locally all at once or hosted somewhere
Off topic, but I assume the name braintrust comes from Creativity, Inc. Amazing book by Pixar co-founder Edwin Catmull.
Redefining commonly understood phrases to mean something else in your own little world make you look ignorant.
Indeed! Why not just call it "asynchronous software development" or something similar? "asynchronous programming" is a bad choice, partly because it will be un-googleable.
The intent is more likely clickbait.
Not that catchy (even in fewer words).When this bubble finally pops, someone is going to have to clean up all the nonsense AI code out there.
Dream on!
i was reading this article while my laptop was auto-vscoding
He says define the problem like its the easy part. If we had full specs, life would be a lot easier.
They don't really as once the spec gets detailed enough it becomes so large and unwieldy that nobody with any actual power reads the things.
An executive at a large company once told me about something where a spec had been written and reviewed by all relevant stakeholders: "That may be what I asked for, but its not what I want."
[dead]
[dead]
Awful font.